Anthropic lets a Claude agent write a program that runs 1,000 agents

Running one agent across hundreds of pieces of work has always meant somebody writing a loop. Anthropic turned on a feature on 9 October 2026 that moves the loop into the agent: a Claude Managed Agents session can now hand a task to a workflow run that writes its own program, spawns the agents it needs, collects what each one returns and ends when the result is combined.

The feature is called dynamic workflows and is in beta behind the header managed-agents-2026-04-01. It is enabled by default in the multiagent_20261001 configuration block. Anthropic frames it as the method for "work with many pieces, parallel work, or a long task that should finish sooner", with audits, migrations, deep research and cross-checking as named examples. The alternative, subagents, still exists for a case where the lead agent needs to send follow-up messages to a specialist thread, and the two can be enabled together.

What the agent is allowed to do, and the limits that bind it

The caps are stated in the Workflow runs documentation. A single run can start up to 1,000 agents over its lifetime, with 64 threads working at once (not guaranteed by the API). Default run lifetime is 24 hours, and the agent can shorten it when it starts the run. A session has 10 open runs at once by default; past that, the next request returns max_workflow_runs_error. If a run passes its lifetime it ends with timeout_error, including when a pause at the session budget held it idle for the whole 24 hours.

The agent writes the workflow itself, as a program that runs other agents in phases and combines the results. Context and results pass programmatically from one agent to another, so the primary thread is free to answer the user while the run goes on in the background. Each agent in a run works in its own session thread with its own history, but every thread shares the session's sandbox, filesystem and vault credentials. The agent can start more than one workflow run on one turn and watch all of them through workflow_run.* events on the primary stream.

Where developers of the parallel loop went

Before this, a team wanting a Claude agent to review 300 documents in parallel wrote the splitter: a function that counted the documents, started subagents, picked up what each one returned and merged the output. The system prompt had to tell the agent when to call it. The new path puts the splitter inside the agent. The developer writes one sentence into the system prompt, such as "when you are asked to review more than a few contracts, start a workflow run that reads them in parallel", and the agent generates the program each time.

Two consequences that Anthropic names on its own documentation page are worth planning for. The first is budget: every agent in a run costs tokens, so a session budget caps the whole workflow and a run paused at the budget keeps burning down its 24-hour lifetime. The second is session threads: a run's threads are exempt from the session's 25 child-thread limit and the server archives them itself at the end of the run, so a developer who was writing their own cleanup loop can stop.

The agent is deciding when to fan out, which means the system prompt now has to teach it what counts as "many pieces". A sentence that is too loose turns a one-off question into a thousand threads, and the error path is a 24-hour run that bills every token on the way to the lifetime timeout.

Source

Multiagent orchestration and Workflow runs in Anthropic's Claude Managed Agents documentation, with the 9 October 2026 entry in the release notes.